Faithfulness-QA: A Counterfactual Entity Substitution Dataset for Training Context-Faithful RAG Models

1. Problem Statement

Retrieval-Augmented Generation (RAG) systems are designed to ground Large Language Models (LLMs) in external evidence. However, a critical failure mode known as unfaithfulness occurs when models ignore the retrieved context and rely instead on their internal parametric memory (knowledge acquired during pre-training).

• Root Cause: Standard next-token prediction losses treat all tokens equally, failing to distinguish between context-derived information and memorized facts. Without explicit training signals that reward context-grounding, models default to internal knowledge when conflicts arise.
• Gap: Existing solutions either focus on inference-time conflict resolution (e.g., Self-RAG, FaithfulRAG) or provide small-scale evaluation benchmarks (e.g., FaithEval with ~4.9K samples). There is a lack of large-scale, high-quality training data specifically designed to force models to prioritize retrieved context over internal memory through controlled knowledge conflicts.

2. Methodology: Faithfulness-QA Construction

The authors propose Faithfulness-QA, a dataset of 99,094 samples constructed via a fully automated, four-stage pipeline. The core mechanism is Counterfactual Entity Substitution, which creates artificial conflicts between the context and the model's likely parametric knowledge.

A. Source Data

The dataset is built from two complementary extractive QA benchmarks:

1. SQuAD: 87,599 pairs (Wikipedia paragraphs, median length 701 chars).
2. TriviaQA: 87,041 pairs (Longer evidence documents, median length 1,581 chars).

B. The Four-Stage Pipeline

1. Entity Bank Construction:

   • Extracted named entities from source contexts using SpaCy's en_core_web_lg NER model.
   • Curated a bank of 76,953 unique entities across 8 types: PERSON, GPE, ORG, DATE, CARDINAL, NORP, LOC, and EVENT.
   • Filtered entities by length (2–100 characters).

2. Answer–Entity Matching:

   • For each QA sample $(q, c, a)$, the system attempts to match the answer $a$ to a recognized NER entity in the context $c$.
   • Uses a three-strategy cascade: Exact match, Substring match, and Positional overlap.
   • Samples failing to match a named entity are filtered out (this is the primary bottleneck).

3. Counterfactual Entity Substitution (Core Step):

   • Identifies the original answer entity ($e_{orig}$) and its type ($t$).
   • Samples a replacement entity ($e_{new}$) from the bank that shares the same type $t$ but is distinct from $e_{orig}$.
   • Enforces length compatibility ($0.3 \le |e_{new}|/|e_{orig}| \le 3.0$).
   • Global Replacement: Replaces all case-insensitive occurrences of $e_{orig}$ with $e_{new}$ in the context to create $c'$.
   • The new "faithful" answer becomes $e_{new}$.

4. Quality Filtering & Splitting:

   • Applies 6 rule-based checks (e.g., ensuring the new entity is in the modified context, the original is removed, context length is sufficient, no excessive replacements).
   • Splits passing samples into Train/Dev/Test (80/10/10) with a fixed seed (42).
   • The entire pipeline is deterministic, runs on a single CPU in ~5.5 hours, and requires no GPU.

3. Key Contributions

1. Faithfulness-QA Dataset: A large-scale (99K samples) resource specifically designed to train RAG models to follow context over parametric memory. It is roughly 20x larger than existing faithfulness evaluation benchmarks.
2. Automated Construction Pipeline: A fully automated, open-source pipeline achieving a 56.7% substitution success rate with 100% quality pass rates on audited samples.
3. Typed Entity Bank: A reusable resource of 76,953 entities across 8 categories, enabling reproducible and extensible dataset construction.
4. Comprehensive Analysis: Detailed statistical analysis of entity distributions, context lengths, and failure modes, demonstrating the complementarity of SQuAD and TriviaQA sources.

4. Results and Statistics

• Scale: 99,094 total samples (49,094 from SQuAD, 50,000 from TriviaQA).
• Success Rate: 56.7% overall (56.0% for SQuAD, 57.4% for TriviaQA).
• Failure Analysis: The primary reason for filtering (64.0%) is the lack of a matching NER entity (e.g., answers are descriptive phrases or boolean values). This is a deliberate trade-off to ensure type-consistent, well-defined substitutions.
• Quality Audit: A random 200-sample audit achieved 100% pass rates across four dimensions: replacement presence, original removal, context modification, and length ratio.
• Entity Distribution:
  • TriviaQA is dominated by PERSON entities (45.7%).
  • SQuAD shows a more uniform distribution across ORG, DATE, and PERSON.
  • The combination ensures broad coverage, mitigating type-specific overfitting.

5. Significance and Applications

Faithfulness-QA addresses a critical gap in the RAG ecosystem by shifting the focus from evaluating faithfulness to training for it.

• Faithfulness-Aware Fine-Tuning: Enables training models to override parametric memory when it conflicts with retrieved evidence.
• Attention-Based Loss Functions: The explicit modification regions allow for supervision of cross-attention weights, encouraging models to attend to the context.
• Knowledge Conflict Research: Provides a controlled environment to study how LLMs resolve conflicts between external evidence and internal beliefs.
• Complementarity: While FaithEval serves as a benchmark for evaluation, Faithfulness-QA serves as the training resource to build models capable of passing such benchmarks.

6. Limitations and Future Work

• Limitations:
  • No coreference resolution (pronouns referring to replaced entities are not updated).
  • Relies on rule-based filtering rather than NLI models for semantic consistency.
  • Some substitutions may be syntactically valid but semantically implausible (e.g., "Amritsar, Illinois").
  • Currently limited to English.
• Future Directions: Integrating coreference resolution, using LLMs for entity matching to improve success rates, applying NLI-based filtering, and conducting downstream training experiments with open-source LLMs (Llama-3, Mistral, Qwen).

Conclusion: Faithfulness-QA provides the necessary infrastructure to move RAG systems from "retrieval-augmented" to "retrieval-faithful," ensuring that models genuinely rely on the evidence provided to them rather than their pre-trained biases.